Looking at your graph again, I think you're losing nearly
10% 5% of the maximum pump power as avalanche energy in the mosfet, except when you are at 100% PWM and the mosfet is always on. I expect the mosfet is getting warm if not hot. You haven't said what the current rating of the pump is, although I see you're using a 4A fuse.
[EDIT: I now think you are likely to be within the repetitive avalanche energy of the mosfet, although adding a flyback diode is kinder to the mosfet and the motor, and should remove the kink in the graph.]
I am not sure I would go as high as 1000uF bulk decoupling I am sure 470uF would be good enough.
Very likely true, however the ripple current rating should be considered too. It would be interesting to see a scope trace of the signal across R4 so that we can tell what the current is doing.
An old trick to get round this was to disconnect the earth lead from the scope and have it poking out of the mains socket to remind you, but in these days of sealed plugs and ground political correctness it is frowned upon.
A modern alternative to this trick is to use one of the USB-powered digital storage scopes that uses a PC as a display (e.g. PicoScope for around GBP300 for a good one, or a similar spec one from Hong Kong for around GBP100), and use it with a laptop running on battery power.